DAMEX: Dataset-aware Mixture-of-Experts for visual understanding of mixture-of-datasets

Authors: Yash Jain, Harkirat Behl, Zsolt Kira, Vibhav Vineet

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on Universal Object-Detection Benchmark show that we outperform the existing state-of-the-art by average +10.2 AP score and improve over our non-Mo E baseline by average +2.0 AP score.
Researcher Affiliation Collaboration Yash Jain1 Harkirat Behl2 Zsolt Kira1 Vibhav Vineet2 1Georgia Institute of Technology 2Microsoft Research
Pseudocode No The paper describes methods using mathematical equations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/jinga-lala/DAMEX.
Open Datasets Yes UODB comprises of 11 datasets: Pascal VOC [5], Wider Face [40], KITTI [8], LISA [26], DOTA [36], COCO [22], Watercolor, Clipart, Comic [13], Kitchen [9] and Deep Lesions [38], shown in Figure 1.
Dataset Splits Yes All the reported numbers in this work are mean Average Precision (AP) scores evaluated on the available test or val set of corresponding dataset.
Hardware Specification Yes We kept one expert per GPU and train on 8 RTX6000 GPUs with a batch-size of 2 per GPU, unless mentioned otherwise.
Software Dependencies No The paper mentions using the 'TUTEL library [12]' but does not provide specific version numbers for this or any other software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes For hyper-parameters, as in DINO, we use a 6-layer Transformer encoder and a 6-layer Transformer decoder and 256 as the hidden feature dimension. We use a capacity factor f of 1.25 and an auxiliary expert-balancing loss weight of 0.1 with top-1 selection of experts. We use a learning-rate of 1.4e-4 and kept other DINO-specific hyperparameters same as [42].